Project

Towards Robust and Automated Transformation Generation for Schema-Mapped Data

Objective

Project Objectives: 1. Dynamic Validation without Ground Truth - Develop methods to validate generated transformations using mined specifications and synthetic test data. - Investigate whether synthetic data and fuzzing can replace explicit user validation. 2. Support for Multi-Pattern Inputs - Detect heterogeneous data patterns in source columns (e.g., different date formats). - Cluster inputs by pattern and synthesize a separate transformation for each cluster. 3. Automatic Multi-Dimensional Validation - Design automatic validators to assess transformations across: - Semantic validity (does the output make sense?), - Data validity (parseability, type checks), - Runtime efficiency (execution cost, code quality), - Readability (documentation, maintainability). Methodology: - Extend the existing LLM-based pipeline that uses sample selection, few-shot prompting, and error-driven repair. - Integrate dynamic validation by generating synthetic test cases and checking semantic/data consistency. - Implement clustering + routing methods for multi-pattern inputs. - Develop lightweight validators for runtime, readability, and correctness. Experiments: - Evaluate on existing public benchmarks (FlashFill, IJCAI-data, SyGuS) and create new ones. - Metrics: transformation accuracy, robustness to heterogeneous inputs, validation precision/recall, runtime cost. - Ablation studies: with vs. without dynamic validation, single vs. multi-pattern transforms, baseline vs. validator-guided pipeline.

Outcome

Publication in a tier 1 AI conference